The present invention relates to the efficient processing of memory requests in cache coherent distributed shared memory multiprocessor systems. More specifically, the present invention leads to improved processing speed and reduced size of directory caches in the coherence controller of shared memory multiprocessor systems by utilizing an efficient directory cache replacement policy.
Conventional computer systems often include on-chip or off-chip cache memories that are used with processors to speed up accesses to system memory. In a shared memory multiprocessor system, more than one processor, sometimes referred to as a compute node, can store a copy of the same memory locations (or lines) in their respective cache memories. An internode cache coherence mechanism is required to maintain consistency among the multiple cached copies of the same memory line across all nodes that store the line. In small, bus-based multiprocessor systems, the internode cache coherence mechanism is usually implemented as a part of the cache controllers using a snoopy coherence protocol. The snoopy protocol cannot be used in large systems that are connected through a non-broadcast interconnection network since snoopy protocols may require all requests to be broadcast to all nodes. As a result, these systems use a directory-based protocol to maintain cache coherence, where a memory directory in each node is associated with the main memory. The coherence controllers of cache coherent distributed shared memory multiprocessor systems use these memory directories to maintain a record of the state of memory lines in the caches of the various processing nodes of the multiprocessor system. This state information includes data indicating which cache(s) has a copy of a given line or whether the line has been modified in a cache(s).
FIG. 1 illustrates a representation of an example of a shared memory multi-node computing system. On a system area network 100, one or more nodes exist. The system can include one or more compute nodes 110, memory nodes 112, and/or compute/memory nodes 114. Each compute node 110 includes one or more processors with associated caches 120, one or more shared/remote caches 130 (optional), and one or more coherence controllers 160. Caches 120 and 130 can be referred to collectively as xe2x80x9cdata cache.xe2x80x9d Each memory node 112 contains one or more main memory modules 140, at least one memory directory 150 which stores information about the contents of its associated main memory 140, at least one coherence controller 160, at least one directory cache (DC) 170, as disclosed in the parent application, and may also include one or more I/O devices (not shown). A compute/memory node 114 includes the elements of a compute node and a memory node.
Memory directories such as 150 may be organized as xe2x80x9cfull mapxe2x80x9d memory directories where the state information on every single memory line is stored by mapping each memory line to a unique location in the directory. That is, there is a one to one state mapping between a main memory line and a memory directory entry. As a result, when the size of main memory increases, the memory directory size must also increase. If the memory directory is implemented as relatively fast static RAM, tracking the size of main memory becomes prohibitively expensive. If the memory directory is implemented using slow static RAM or DRAM, higher cost is avoided, but a penalty is incurred in overall system performance due to the slower chips. In fact, each directory access in such slower implementations will take approximately 5-20 controller cycles to complete.
In order to overcome the problems with full map memory directories, xe2x80x9csparsexe2x80x9d memory directories have been used instead. In a sparse memory directory multiple memory lines are mapped to a single location in the sparse directory, although state information stored in the mapped location may be valid only for a subset of the memory lines mapping to that location. Due to its smaller size, a sparse directory can be implemented in an economical fashion using fast static RAMs. However, when there is contention among memory lines for the same sparse directory entry field, i.e., when it is desired to store directory information for a given line of memory in a location that already contains current directory information for another line of memory, the state information of the line already stored in the memory directory must be replaced. Since there is no backup state information, when a line is replaced in the sparse directory, all the caches in the overall system having a copy of that line must be asked to invalidate their copies. This incomplete directory information leads to both coherence protocol complexity and performance loss.
Since directory lookup speed and throughput have been shown to have an important impact on the overall performance of distributed shared memory multiprocessor systems, directory caches, as disclosed in the parent application, having records corresponding to a subset of the memory directory entries can be used. Such smaller and faster directory caches associated with the coherence controllers can be used to improve the speed of directory lookup by keeping the copies of a subset of the memory directory entries in faster on-chip or off-chip memory.
FIG. 2 shows a representation of a directory cache (xe2x80x9cDCxe2x80x9d) system 200 which caches the state information for a subset of memory lines 250 of the main memory 220. While, in this representation, the memory directory 210 is illustrated as a full map memory directory, the memory directory could also be implemented as a sparse directory, or other directory scheme. The DC 200 is organized into DC lines 230 and stores state information from only a subset of the directory entries 240.
FIG. 3 shows an organization of a DC, such as DC 200 of FIG. 2. The DC 200 stores s sets 310. Each set consists of w DC lines 320 corresponding to w ways of set associativity. Each DC line 320 contains a validity indicator 330, a tag 340, and e memory directory entries 350 each containing state information 360 for a line of the associated memory, where w and e are integers greater than or equal to 1 and s is typically a power of 2. The validity indicator 330, which can be a single bit or multiple bits, indicates that the DC line currently holds valid information. The tag 340 maps the directory entries 350 in the cache line back to their locations in memory directory 210.
Improved performance for the DC, and accordingly for the entire multiprocessor system, can be achieved if the subset of memory directory entries cached in the DC consists of the most frequently used entries; that is when the hit ratio of the DC is high. The hit ratio of a cache is mainly dependent on the allocation and replacement policies of such a cache.
In conventional caches, one widely used allocation policy caches each and every memory line (or each memory directory entry or set of entries in the case of directory caches) that is referenced and misses in the cache. A less common policy is to designate parts of memory (or the memory directory in the case of directory caches) as uncacheable, so that uncacheable lines do not compete for limited cache space (a.k.a. cache pollution) with cacheable lines that are more likely to benefit from caching, as judged by the programmer. The problem with this latter policy, if applied to directory caches, is that it requires operating system support and programmer annotations, which are undesirable as they require significant extra programmer effort and, at times, may be impossible to implement.
Two common cache replacement policies are Least Recently Used (LRU) and random replacement. LRU is based on the heuristic that memory lines (or memory directory entries in the case of directory caches) that have not been used recently are less likely to be accessed in the near future than others. The random policy, as implied by its name, replaces entries randomly in cases of conflict.
In distributed shared memory multiprocessor systems, only a small part of what is designated by the operating system as shared memory is actually shared between multiple nodes. Most lines designated as shared memory lines are actually accessed by only one processor on its local node. Using conventional allocation and replacement policies typically results in the pollution of directory caches with directory entries of cache lines that are not in fact shared and that typically remain cached in the conventional caches of local compute nodes after their initial access without causing further accesses to the coherence controller. Conventional caching policies fail in this regard, and therefore methods that adequately address this issue for directory caches are needed.
It is an object of the invention to provide a cache system that minimizes directory cache pollution, leading to higher directory cache hit ratios and resulting improvement in system performance.
It is another object of the invention to reduce directory cache pollution by providing a directory cache allocation method that utilizes memory line state information.
It is another object of the invention to reduce directory cache pollution by providing a directory cache eviction method that utilizes memory line state information.
It is another object of the invention to improve the quality of memory line state information by including a memory line""s sharing history information.
It is another object of the invention to identify lines of shared main memory that are not in fact shared by other nodes in a system and not allocate the directory entries corresponding to such lines in the directory cache, and to evict DC entries representing lines of shared main memory that are never or seldom accessed by more than one node, in order to minimize DC pollution and thereby increase hit ratio, improve the performance of the coherence controllers and, consequently, the overall system performance.
In accordance with one aspect of the invention, the foregoing objectives are realized by providing a system for maintaining consistent cached copies of memory in a multi-node computing system having a main memory, comprising at least one memory directory having memory directory entries mapping the main memory, one or more of the memory directory entries including state information for a corresponding line of main memory, at least one directory cache for storing directory cache lines corresponding to a subset of the memory directory entries, and means for using the state information to allocate memory directory entries to the directory cache.
A second aspect of the invention provides a system for maintaining consistent cached copies of memory in a multi-node computing system having a main memory, comprising at least one memory directory having memory directory entries mapping the main memory, one or more of the memory directory entries including state information for a corresponding line of main memory, at least one directory cache for storing directory cache lines corresponding to a subset of the memory directory entries, and means for using the state information to evict directory cache lines in the directory cache.
A third aspect of the invention provides an apparatus for maintaining consistent cached copies of memory lines in a multi-node computing system having a main memory, comprising at least one memory directory having memory directory entries mapping the main memory, and memory directory entries, stored in the memory directory, including state information indicating whether one or more nodes in the multi-node computing system have an associated data cache that contains a copy of a memory line corresponding to the memory directory entry.
A fourth aspect of the invention provides a method of maintaining state information in a memory directory entry of a multi-node computing system having a main memory, one or more nodes in the system having a coherence controller, a processor cache, a memory directory of the main memory and a directory cache of the memory directory, the method comprising the steps of receiving a signal at the directory cache in one node of the system indicative of a coherence request for a cached memory line from one of the other nodes of the system, performing a memory directory lookup to determine the location of the directory entry of the cached memory line, and storing information describing sharing behavior of the cached memory line.
A fifth aspect of the invention provides a method of directory cache eviction in a multi-node computing system having a main memory, one or more of the nodes having a coherence controller, a processor cache, a memory directory associated with the main memory, and a directory cache of the memory directory having directory entries storing state information for associated lines of main memory, the method comprising the step of selecting a directory cache line for eviction based on the state information of the retrieved directory cache line, and updating a validity indicator associated with the selected cache line to indicate that the directory cache line does not include valid information.
A sixth aspect of the invention provides a method of directory cache allocation in a multi-node computing system having a main memory, one or more of the nodes in the system having a coherence controller associated with the main memory, memory directory associated with the main memory, a directory cache of the memory directory having directory entries storing state information for associated lines of main memory, the method comprising the steps of retrieving a memory directory entry from a memory directory, and deciding whether to allocate the memory directory entry based on the state information of the memory directory entry.